Researchers have introduced UFO, a unified framework for evaluating omni-condition alignment in multi-modal image generation, particularly for subject-driven customization. The work addresses a gap in evaluation methods, which have lagged behind rapid advances in generative models.
Existing evaluation approaches—whether embedding-based or based on Multi-modal Large Language Models (MLLMs)—assess alignment with each modal condition in isolation. According to the paper, this contradicts the simultaneous condition alignment objective of multi-modal image generation and leads to poor consistency with human judgments.
UFO proposes an Atomized Chain-of-Evaluation paradigm. It decomposes omni-condition alignment into a sequential chain of fine-grained, disentangled Atomic Evaluation Units (AEUs), categorizes them into distinct modality-relevance classes, and then uses general or dedicated functional calls to verify different AEU types accurately.
Experimental results show that UFO achieves the highest correlation with human evaluation preferences, with an average improvement of 15.25%. The authors also present UFO-Bench, a dedicated benchmark designed to holistically evaluate existing customization models under diverse mutual interactions of textual and visual conditions.